🔍 Read the full analysis: Why Readers Tire Of AI Text: Six Patterns And Ways To Improve Them on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A preprint published October 21 compares writing from 14 language models with human text and reports six language patterns used more often by larger, advanced models. The authors say editing for clarity may help, but they did not test comprehension with readers, and the study has not been peer reviewed.
A new linguistics preprint reports that text from 14 AI models used six sentence patterns more often than comparable human writing, offering one possible explanation for why some chatbot answers can feel harder to read. The study, posted to arXiv on October 21, compares model output across chat, coding and persuasive-writing tasks, but does not test whether readers actually understand the AI text less well.
Authors Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin examined text from 14 models. Their material included 120 simulated chats, 45 coding sessions based on GitHub projects and 980 short persuasive essays. Nine models were classed by the authors as frontier or near-frontier; five smaller open models came from the Llama and Olmo families.
The researchers compared the outputs with human writing matched to the task, including subtitles, StackExchange posts, student essays and Wikipedia. They also translated sentences into Japanese and back into English as a control: the authors reasoned that if a pattern became less common after the wording was altered while meaning remained, the model had selected it despite a plainer alternative being available.
The six patterns were unusual word pairings, long distances between words that belong together, omission of the optional word “that,” strings of three or more nouns, nouns repurposed as verbs, and descriptions in which inanimate objects act like living things. According to the source account of the paper, all nine larger models used more rare word pairings and more personification than human writers. Smaller models with lighter post-training were described as writing more like people. Those findings describe patterns in this sample, not every AI response.
The study at a glance
14 models, 3 settings, 6 patterns
Three linguists compared AI text with human text of a similar kind, then checked which hard-to-read patterns the models chose on their own.
The models
Frontier and near-frontier models are highlighted.
The comparison
- Chat vs. film and TV subtitles
- Code vs. StackExchange posts
- Essays vs. 2,934 high-scoring student essays (grades 6 to 12) and Wikipedia
- Control: every sentence translated into Japanese and back. If a pattern gets rarer, the model chose it over a plainer option.
- Cost: $251.19 in API fees for the whole data collection.
Held for all nine
More rare word pairs and more objects that act like living things than human writers.
Grew with training
In the Olmo family the effects rose from base to instruct to the reasoning model, Olmo Think.
Fixable
Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty comparisons.
Source: Clark, Martin & Mahowald, Why is AI-generated language so hard to understand?, arXiv preprint, 21 October 2026 (not peer reviewed). Cost figure from the paper’s appendix A.6.
Powered by Thorsten Meyer AI
Why These Patterns Slow Reading
The study gives editors and AI users a set of specific features to look for instead of treating “AI-sounding” prose as a single, vague problem. A phrase such as a noun stack can force readers to work out how several concepts relate; a long gap between a subject and its verb can make them hold information in memory. The researchers draw on earlier studies that have linked such features to slower processing.
That distinction matters for teams using generated text in instructions, customer communications or other material readers need to understand quickly. The source article recommends editing for clarity—shortening sentences, unpacking noun strings and choosing familiar phrasing—rather than relying on an AI detector. It also notes that detectors have a weak record and can incorrectly flag human writing. The present preprint, however, does not itself establish that its six features reliably identify AI or that removing them improves reader comprehension.
The authors suggest that training incentives may help explain the patterns: systems optimized for mathematics, coding and reasoning may be rewarded for dense answers rather than easy reading. That is the researchers’ proposed explanation, not a cause demonstrated by the comparison. Their test involving a smaller model is described as a proof of concept, not conclusive evidence about training effects.
Products Worth Considering
Samsung 32-Inch Odyssey OLED G8 (G80SH) 4K Gaming Monitor
Bose QuietComfort Ultra Earbuds (2nd Gen), Deep Plum
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Portable Waterproof Speaker, Black
As an affiliate, we earn on qualifying purchases.
How Researchers Compared the Models
The paper, titled “Why is AI-generated language so hard to understand?”, is an arXiv preprint, meaning it has been shared publicly before peer review. The authors drew on several forms of model output and matched them with different human-writing sources, rather than comparing every task against a single general corpus. The source article says the data collection cost $251.19 in API fees, according to the paper’s appendix.
The source account gives examples of the patterns, but distinguishes its own illustrations from quotations in the research paper. One paper example from Claude Sonnet 5 reads: “The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.” The article’s suggested edits are practical examples, not experimentally validated fixes. The study is limited to English-language material; its authors did not report a test of real readers’ comprehension.
“The guitar sitting in the corner because you’re still ‘deciding’ teaches you zero chords.”
— Claude Sonnet 5, quoted in the paper’s Table 2
What the Preprint Cannot Establish
The findings have not yet been peer reviewed, and they cover English text from a particular set of models and tasks. The source material does not provide enough detail to independently assess how representative the human comparison texts are or whether the same six patterns appear at similar rates in other languages, genres or model versions.
Most importantly, the authors did not test whether people reading the sampled passages understood them less accurately or more slowly. Earlier research cited by the authors links some features to processing difficulty, but that is not the same as showing that this study’s model outputs caused a measurable comprehension problem. The claim that training for coding, mathematics and reasoning encourages dense prose is also a proposed explanation, not a confirmed finding.
The source material ends partway through its discussion of personification and does not give full numerical results for each pattern. It also does not specify a planned peer-review date or follow-up study. Readers should treat the reported comparisons as preliminary evidence, not a universal checklist for detecting machine-written text.
Further Testing and Editorial Use
The next test would be to measure how readers respond to matched passages: whether they read them more slowly, misunderstand them more often or prefer a revised version. Independent replication across languages, model families and writing tasks could show whether the six patterns hold beyond this study’s sample. Peer review may also examine the methods and strength of the researchers’ interpretations.
For now, writers can use the patterns as an editing checklist without assuming that any one feature proves a passage was generated by AI. Put subjects close to their verbs, clarify compressed noun phrases and replace strained wording with familiar language. Whether those revisions produce measurable gains for readers remains to be tested.
Key Questions
What did the study examine?
It compared text from 14 language models with human writing across simulated chats, coding sessions and persuasive essays, looking for six language patterns.
What are the six patterns?
They are rare word combinations, long gaps between related words, omitted “that,” noun strings of three or more words, nouns used as verbs and inanimate things described as if alive.
Does the study prove AI text is harder to understand?
No. The authors did not test real readers’ comprehension of the sampled passages. The preprint connects its features to earlier research on processing difficulty, but the study does not directly establish a comprehension effect.
Can these patterns identify text written by AI?
Not reliably on the evidence provided. The study reports differences between sampled model and human writing; it does not validate a detector. The source article advises editing for clarity rather than using detectors as a substitute.
Has the research been peer reviewed?
No peer review is reported in the source material. The paper was posted to arXiv on October 21, 2026 as a preprint.
Source: ThorstenMeyerAI.com
